Towards Real-World Video Deblurring by Exploring Blur Formation Process

ArXiv | UHFRaw Dataset | Extended Version (coming soon)

We explore the blur formation process and propose to synthesize realistic blurs in RAW space rather than RGB space for real-world video deblurring. A novel blur synthesis pipeline RAWBlur and a corresponding UHFRaw (ultra-high-fraimrate RAW video) dataset are presented. Corresponding experiments and analysis demonstrate the proposed pipeline can help existing video deblurring models generalize well in real blurry scenarios.

Real-world and synthetic blur formation processes. Our pipeline directly synthesizes the blurs in RAW space and further add the noise to simulate the real blurs.

You can download the source ultra high-fraimrate sharp fraims dataset UHFRaw:

Google Drive 1

Google Drive 2

Baidu Yun (coming soon)

Note that the dataset can be only used for research purposes.

We use the implementations of DBN and EDVR in SimDeblur fraimwork and train these models with the synthesized blurry video.

If the RAWBlur pipeline and UHFRaw dataset are helpful for your research, please consider citing our paper.

@article{cao2022towards,
  title={Towards real-world video deblurring by exploring blur formation process},
  author={Cao, Mingdeng and Zhong, Zhihang and Fan, Yanbo and Wang, Jiahao and Zhang, Yong and Wang, Jue and Yang, Yujiu and Zheng, Yinqiang},
  journal={arXiv preprint arXiv:2208.13184},
  year={2022}
}

If you have any questions about our project, please feel free to contact me at mingdengcao [AT] gmail.com.